1 Definition and basic notions

1.1 Integrable families and the role of L¹

Uniform integrability (UI) is defined for a collection of integrable random variables \(\{X_\alpha\}\) on a probability space \((\Omega,\mathcal{F},\mathbb{P})\). The central setting is that each \(X_\alpha\) belongs to \(L^{1}(\mathbb{P})\), meaning \(\mathbb{E}X_\alpha<\infty\). UI strengthens mere integrability by requiring that the family does not “concentrate mass” farther and farther out in the tail as \(\alpha\) varies.

1.2 Uniform integrability via tail control

A standard definition uses tail integrals. A family \(\{X_\alpha\}\subset L^{1}(\mathbb{P})\) is uniformly integrable if for every \(\varepsilon>0\) there exists \(M>0\) such that \[

\sup_{\alpha}\mathbb{E}\bigl(X_\alpha\mathbf{1}_{\{X_\alpha>M\}}\bigr)<\varepsilon.

\] Intuitively, once values are truncated above a sufficiently large threshold, the remaining expected contribution from the tail becomes uniformly negligible across the entire family.

1.3 Equivalent characterizations

Several equivalent formulations are commonly used. One replaces the indicator truncation by truncated expectations: \[

\lim_{M\to\infty}\sup_{\alpha}\mathbb{E}\bigl(X_\alpha\,;X_\alpha>M\bigr)=0,

\] which is the same tail condition written with semicolon notation. Another viewpoint expresses UI through controlling integrals over sets of small probability: the family is UI if for every \(\varepsilon>0\) there exists \(\delta>0\) such that for all measurable \(A\) with \(\mathbb{P}(A)<\delta\), \[

\sup_\alpha \mathbb{E}\bigl(X_\alpha\mathbf{1}_A\bigr)<\varepsilon.

\] This “small set” criterion is often useful when convergence is described in terms of events rather than thresholds.

1.4 Relationship to absolute uniform integrability

In many settings one considers both \(X_\alpha\) and \(X_\alpha\). Uniform integrability is typically defined for \(X_\alpha\) or equivalently for the absolute values. This leads to the notion of “absolute UI,” where \(\{X_\alpha\}\) is UI in the tail sense. For integrable families, working with absolute values avoids sign issues and provides cleaner stability under dominated convergence arguments.

2 Uniform integrability in probability spaces

2.1 Random variables versus measurable functions

UI is formulated for measurable functions on a probability space, which includes both random variables and more general measurable maps. If the probability space is fixed, UI is an intrinsic property of the measurable functions, not of any particular random-variable representation. Measurability ensures that truncations and set-based integrals used in definitions are well defined.

2.2 Convergence modes compatible with uniform integrability

UI is most valuable when paired with a convergence mode. Typical hypotheses are:

  • convergence in probability,
  • almost sure convergence,
  • or convergence in distribution together with additional structural assumptions.

The key idea is that UI allows expectations to follow the limiting behavior of the variables. Without UI, convergence in probability may occur while expected values fail to converge because rare events carry increasing weight.

2.3 Tightness and moment conditions

In many applications, UI is connected to tail control and, consequently, to moment bounds. For instance, uniform boundedness of an \(L^{p}\) norm with \(p&gt;1\) implies UI in \(L^{1}\) via Hölder’s inequality. More generally, UI can be viewed as a form of “tightness in expectation”: it prevents escape to large values and keeps the contribution of extreme outcomes under control.

2.4 Examples and counterexamples

Example (bounded in \(L^p\) for \(p>1\)). If \(\sup_\alpha \mathbb{E}X_\alpha^{p}<\infty\) for some \(p>1\), then \(\{X_\alpha\}\) is uniformly integrable. A tail estimate of Chebyshev/Markov type yields the required vanishing of \(\mathbb{E}(X_\alpha\mathbf{1}_{X_\alpha>M})\) uniformly in \(\alpha\).

Example (failure of UI). Let \(X_n = n\mathbf{1}_{\{U\le 1/n\}}\) where \(U\sim \text{Uniform}(0,1)\). Then \(\mathbb{E}X_n=1\) for all \(n\), but the mass is pushed to increasingly rare events with increasingly large values. One can show that the tails do not vanish uniformly, so the sequence is not uniformly integrable.

These contrasting examples clarify that integrability of each \(X_n\) is not enough; the family must resist systematic tail growth.

3 Fundamental theorems involving uniform integrability

3.1 Vitali convergence theorem (uniform integrability + convergence)

A central result is Vitali’s theorem: if \(X_n\to X\) in probability and \(\{X_n\}\) is uniformly integrable, then \(X\) is integrable and \[ \mathbb{E}X_n \to \mathbb{E}X. \] Variants exist for almost sure convergence with UI as well. The theorem is a probabilistic analogue of convergence principles in measure theory: UI bridges the gap between pointwise convergence and convergence of integrals.

3.2 Convergence of expectations under uniform integrability

More generally, UI permits transferring convergence of random variables to convergence of their expectations for \(L^{1}\)-convergent limits. If \(X_n\to X\) almost surely (or in probability) and \(\{X_n\}\) is UI, then \(\mathbb{E}X_n-X\to 0\) in favorable situations, and at minimum \(\mathbb{E}X_n\to \mathbb{E}X\). The precise strength depends on the additional convergence and the structure of the family, but the expectation convergence is the canonical consequence.

3.3 Dominated convergence as a special case

If \(X_n\le Y\) for all \(n\) with \(Y\in L^{1}\), then the family \(\{X_n\}\) is uniformly integrable. This is because the tail of \(X_n\) is dominated by the tail of \(Y\):

\[

\mathbb{E}(X_n\mathbf{1}_{X_n>M})\le \mathbb{E}(Y\mathbf{1}_{Y>M})\to 0.

\] Thus the classical dominated convergence theorem fits within the UI framework as a special case where a single integrable dominator provides uniform tail control.

3.4 Scheffé-type relationships and refinements (where applicable)

In nonnegative contexts, additional refinements relate convergence of expectations to convergence in \(L^{1}\). For sequences of densities \(\{f_n\}\) with \(\int f_n=1\) and \(f_n\to f\) almost everywhere, Scheffé’s lemma implies \(\|f_n-f\|_1\to 0\). This frequently complements UI discussions: such density sequences automatically exhibit UI due to normalization and pointwise convergence, leading to strong integral convergence outcomes.

4 Stability properties

4.1 Closure under linear combinations

Uniform integrability is stable under linear operations. If \(\{X_\alpha\}\) and \(\{Y_\alpha\}\) are UI and \(a,b\in\mathbb{R}\), then \(\{aX_\alpha+bY_\alpha\}\) is UI. This follows from subadditivity and the fact that tails of a linear combination can be controlled by tails of the components.

4.2 Convexity and permanence under domination

UI is preserved under convex combinations: if \(\{X_\alpha\}\) is UI, then any family formed from convex mixtures of elements of \(\{X_\alpha\}\) remains UI. Additionally, if \(Y_\alpha\leX_\alpha\) and \(\{X_\alpha\}\) is UI, then \(\{Y_\alpha\}\) is UI as well, since tail contributions shrink under domination.

4.3 Behavior under truncation

Truncating a uniformly integrable family in magnitude produces a controlled remainder. If \(\{X_\alpha\}\) is UI and \(X_\alpha^{(M)}:=X_\alpha\mathbf{1}_{\{X_\alpha\le M\}}\), then \(\{X_\alpha^{(M)}\}\) is uniformly integrable for each fixed \(M\), and the difference \(X_\alpha-X_\alpha^{(M)}\) has uniformly small \(L^{1}\) mass for large \(M\). Truncation is a key technical device in proofs involving UI.

4.4 Products and transformations: conditions and limits

UI under nonlinear transformations depends on growth conditions. Common patterns include:

  • If \(g\) is Lipschitz with \(g(0)=0\), then UI of \(\{X_\alpha\}\) typically yields UI of \(\{g(X_\alpha)\}\).
  • For power-type transforms, additional moment assumptions are usually required to keep tails controlled.
  • For products \(X_\alpha Y_\alpha\), one often needs joint integrability and a mechanism such as UI for one factor together with boundedness (in \(L^p\) or UI) for the other.

These statements are not uniform across all transformations; they depend on how the map amplifies large values.

5 Uniform integrability of sequences and nets

5.1 Criteria for sequences

For sequences \(\{X_n\}\), the defining tail criterion can be tested using a single index parameter \(n\). Many theorems are stated for sequences because they match classical convergence concepts. Nevertheless, UI’s definition does not rely on discreteness of indices; it extends naturally to general directed sets.

5.2 Uniform integrability for families indexed by general sets

For a family \(\{X_\alpha\}_{\alpha\in A}\), UI requires the supremum over \(\alpha\) in the tail condition. This generality is helpful in functional analytic arguments and in proofs where one considers approximations along different nets (for example, in measure-theoretic limits).

5.3 Extraction of uniformly integrable subsequences

If \(\{X_n\}\) is not uniformly integrable, it may still be possible to find a subsequence that is UI under certain assumptions, but this is not automatic. When tightness-like conditions hold and the family is relatively compact in a weak sense, subsequences can inherit UI from limiting behavior. In practice, many UI arguments proceed by verifying UI directly rather than extracting it implicitly.

5.4 Comparison across probability measures (change of measure context)

Under a change of measure, UI can be affected because the tails are weighted differently. If \(d\mathbb{Q}/d\mathbb{P}\) satisfies suitable integrability (e.g., belongs to \(L^{p}\) with conjugate exponent \(q\)), then UI under \(\mathbb{P}\) may imply UI under \(\mathbb{Q}\) for transformed variables. The precise conditions resemble those in Hölder-type estimates and depend on how Radon–Nikodym derivatives interact with the truncation events.

6 Functional-analytic viewpoints

6.1 UI as a condition in Banach function spaces

From a functional-analytic perspective, UI is closely tied to the structure of \(L^{1}\) as a Banach space. A UI family behaves like a set that is relatively compact in the weak topology of \(L^{1}\), at least in a way compatible with limit interchange. This viewpoint emphasizes that UI is not merely a probabilistic convenience but reflects properties of integrable functions as elements of a normed space.

6.2 Connections with weak compactness in L¹

In \(L^{1}\), uniform integrability is often used to establish weak compactness or weak sequential compactness. While the exact characterization depends on the underlying measure space and the form of compactness, UI provides the missing ingredient that distinguishes “bounded in \(L^{1}\)” from “compact enough to pass to limits in expectation.”

6.3 de la Vallée-Poussin criterion and Orlicz-type perspectives

A major structural characterization is de la Vallée-Poussin’s criterion: a family \(\{X_\alpha\}\subset L^{1}\) is uniformly integrable if and only if there exists a convex, increasing function \(\Phi:[0,\infty)\to[0,\infty)\) with \(\Phi(t)/t\to\infty\) as \(t\to\infty\) such that \[

\sup_\alpha \mathbb{E}\,\Phi(X_\alpha)<\infty.

\] This criterion converts UI into an Orlicz-type integrability condition, providing an efficient route to verification using growth functions beyond linear ones.

6.4 Duality and implications for operators

UI interacts with duality in \(L^{1}\) and with operator theory. For instance, sequences or families that are UI can yield convergence results when paired against test functions in the dual space \(L^\infty\). Moreover, bounded linear operators mapping into \(L^{1}\) may preserve UI under appropriate conditions, reflecting how truncation control transfers through boundedness and adjoint relationships.

7 Applications

7.1 Martingales and uniform integrability of families of terminal variables

In martingale theory, UI frequently appears as a condition ensuring convergence of the martingale in \(L^{1}\) and preservation of expectation across the limit. A common pattern is that if a martingale is uniformly integrable, then it converges almost surely and in \(L^{1}\), and the limit variable has expectation equal to the initial value. UI thus serves as the bridge between almost sure convergence and convergence of expected values.

7.2 Interchanging limits and expectations in stochastic settings

UI’s main practical role is to justify exchanging a limiting operation (such as \(n\to\infty\) limits associated with approximation schemes) with expectation. This is relevant whenever stochastic processes or random quantities are approximated and one needs the expectation of the limit rather than only the limit of the expectations. UI supplies the technical justification.

7.3 Statistical learning and risk convergence (expectation stability)

In statistical learning, one often considers random variables representing losses or risks under data-driven procedures. If a learning procedure yields convergence of the loss in probability and the corresponding loss variables are uniformly integrable, then expected risk converges as well. This supports generalization-type statements where one wants expected performance to track the limiting behavior of the algorithm.

7.4 Reliability/queueing: controlling tail contributions (non-technical intuition)

In reliability and queueing theory, quantities like waiting times may have heavy tails. UI-like arguments ensure that when approximating a system or studying a limit regime, the probability of extreme delays does not dominate the expected delay. From a non-technical perspective, UI prevents rare but very large outcomes from overturning average performance estimates in the limit.

8 Computation and verification techniques

8.1 Using de la Vallée-Poussin functions

de la Vallée-Poussin’s criterion is often the most direct verification method. One selects a convex growth function \(\Phi\) that grows faster than linearly and then checks a uniform bound on \(\mathbb{E}\Phi(X_\alpha)\). If such a \(\Phi\) exists, UI follows without needing to compute tail integrals for each threshold.

8.2 Checking tail bounds and truncation integrals

Another approach is to estimate directly \[

\mathbb{E}\bigl(X_\alpha\mathbf{1}_{X_\alpha>M}\bigr).

\] Tail bounds such as Markov or Chebyshev inequalities can reduce the problem to controlling moments or exponential-type tails. Truncation integrals are useful when explicit distributions or tail estimates are available.

8.3 Moment-based sufficient conditions

Uniform bounds on moments provide common sufficient conditions. For example, if \(\sup_\alpha \mathbb{E}X_\alpha^{p}<\infty\) for some \(p>1\), UI follows. Similar statements can be made using mixed or conditional moments, provided they yield uniform control of tails in the \(L^{1}\) sense. These results are frequently used because moment bounds are often easier to establish than full UI.

8.4 Practical pitfalls in verifying UI

Common verification issues include:

- assuming that boundedness of \(\mathbb{E}X_n\) alone implies UI (it does not);
  • overlooking that UI must hold uniformly over the entire family, not merely along a subsequence;
  • relying on convergence without tail control, which can lead to correct pointwise or probabilistic convergence but incorrect expectation limits;
  • using moment arguments with an exponent that is too small (e.g., \(p=1\) does not provide the same tail decay).

Careful checking of uniform tail behavior is therefore essential when applying UI-based theorems.